Toward Equitable Low-Carbon Mobility: Fairness-Aware Demand Prediction for Expanding Bike-Sharing Systems

Man Luo; Yixuan Zhao

cs.LGAug 26, 2026

Toward Equitable Low-Carbon Mobility: Fairness-Aware Demand Prediction for Expanding Bike-Sharing Systems

Authors: Man Luo, Yixuan Zhao

Organizations: Department of Computer Science, University of Exeter, Exeter EX4 4QF, U.K.

Abstract

Bike-sharing systems are an important component of low-carbon urban mobility, but continued expansion creates challenges in both cold-start prediction and equitable resource allocation. Newly deployed stations lack historical ridership records, causing a mismatch between training and inference for graph-based models on evolving networks. Historical demand may also encode structural inequalities, as lower ridership in low-income neighborhoods can reflect limited infrastructure access rather than weak latent demand. Models trained directly on such data may therefore reinforce existing mobility disparities. We propose FairGIN, a fairness-aware graph neural network for demand prediction in expanding bike-sharing systems. FairGIN integrates three components. Expansion-Simulated Increment Training stochastically simulates network expansion during training to reduce the cold-start distribution gap. Attention-Based Knowledge Transfer combines station-adaptive temperature scaling with orthogonal embedding alignment to transfer representations from data-rich existing stations to data-sparse new stations. Fairness-Aware Optimization introduces income-stratified regularization and an equity-calibrated deployment score to support more inclusive station placement. Experiments on NYC and Seattle demonstrate that FairGIN achieves state-of-the-art predictive accuracy across diverse expansion scenarios while substantially reducing income-based disparities without compromising overall system efficiency.

Figures & tables

Panel of Figure 1 — Fig. 1 : Bike-sharing station coverage and neighbourhood income in Manhattan. (a) Spatial distribution of existing stations in 2018 and newly deployed stations from 2018. (b) Per-capita station density across neighbourhood income quintiles.

Fig. 2 : Overview of the proposed FairGIN framework for fairness-aware demand prediction in expanding bike-sharing networks. The framework first constructs a heterogeneous station graph from multi-source station features, then applies expansion-simulated increment training (ESIT) to simulate cold-start expansion scenarios, and finally performs attention-based knowledge transfer with fairness-aware optimization for new-station demand prediction.

	NYC Citi Bike	Seattle Bikeshare
Existing stations ( $N$ )	746	186
New stations ( $M$ )	312	61
Time span	2018 to 2023	2017 to 2018
Total trips	80M+	1.6M+
Temporal resolution	Hourly	Hourly
G + / G - (new)	149 / 163	29 / 32

TABLE I : Dataset statistics.

	NYC Citi Bike					Seattle Bikeshare
Method	MAE $\downarrow$	RMSE $\downarrow$	RFG $\downarrow$	IFG $\downarrow$	$\|\rho\|$ $\downarrow$	MAE $\downarrow$	RMSE $\downarrow$	RFG $\downarrow$	IFG $\downarrow$	$\|\rho\|$ $\downarrow$
ARIMA	4.67 ±.24	7.53 ±.37	2.47 ±.08	2.04 ±.07	0.433 ±.016	3.89 ±.21	6.24 ±.32	2.08 ±.07	1.72 ±.06	0.414 ±.015
LSTM	4.23 ±.18	6.87 ±.24	2.34 ±.07	1.92 ±.06	0.412 ±.014	3.47 ±.14	5.61 ±.19	1.98 ±.06	1.67 ±.05	0.387 ±.013
STGCN	3.81 ±.15	6.12 ±.21	2.28 ±.06	1.87 ±.06	0.398 ±.013	3.12 ±.12	5.08 ±.18	1.93 ±.06	1.62 ±.05	0.371 ±.012
DCRNN	3.64 ±.14	5.93 ±.20	2.31 ±.07	1.89 ±.06	0.403 ±.013	2.98 ±.12	4.87 ±.17	1.96 ±.06	1.64 ±.05	0.378 ±.012
FairST	3.73 ±.14	6.03 ±.20	1.71 ±.05	1.42 ±.04	0.315 ±.010	3.06 ±.12	4.95 ±.17	1.38 ±.04	1.16 ±.04	0.297 ±.010

TABLE II : Prediction accuracy and fairness comparison on new stations (

\mathcal{V}_{B}

). Bold : best; underline : best baseline per metric.

Panel of Figure 5 — Fig. 3 : Performance comparison of ablation variants evaluated by MAE and RFG. (a) NYC. (b) Seattle.

Panel of Figure 6 — Fig. 4 : (a) Sensitivity of MAE and RFG to the ESIT mask ratio $\rho$ . (b) Sensitivity of MAE and RFG to the fairness regularization weight $\lambda_{\mathrm{fair}}$ . Spatial distribution of newly added stations under different expansion patterns: (c) localized expansion, (d) regional expansion. Prediction performance of different methods in terms of MAE and RFG under (e) localized expansion and (f) regional expansion. Prediction performance of different methods across expansion rates on NYC Citi Bike, (g) MAE, (h) RFG.

Dataset	Strategy	Top-10	Top-20	Top-30	Mean G -
NYC	Demand-only	0.20	0.25	0.28	1.84
	Random	0.47	0.46	0.48	2.31
	FairGIN	0.60	0.55	0.53	2.47
	Base rate	0.52			–
Seattle	Demand-only	0.18	0.22	0.27	1.61
	Random	0.49	0.51	0.48	2.04

TABLE III : Deployment equity simulation. G - ratio denotes the share of selected stations in

\mathcal{G}^{-}

Model	Cold	#Params	Train (ms)	Infer (ms)
ARIMA	✗	$\sim$ 5/stn	—	—
UrbanGPT	$\sim$ ✓	$\sim$ 7.2B	—	—
LSTM	✗	200.7K	$1.8_{\pm 1.0}$	$0.29_{\pm 0.02}$
STGCN	✗	59.0K	$3.5_{\pm 1.1}$	$0.95_{\pm 0.19}$
DCRNN	✗	148.7K	$91.4_{\pm 17.2}$	$51.2_{\pm 9.4}$
GraphSAGE	✓	36.7K	$1.6_{\pm 0.1}$	$0.41_{\pm 0.26}$

TABLE IV: Complexity comparison (

N\!=\!37

M\!=\!16

|\mathcal{E}|\!=\!604

; RTX 4090). ✓/✗: cold-start support. Mean

\pm

std (100 inference / 20 training runs).

Panel of Figure 9 — Fig. 5 : Deployment outcomes on the NYC Citi Bike development split with $M\!=\!16$ . Red denotes $\mathcal{G}^{-}$ and blue denotes $\mathcal{G}^{+}$ . Panel (a) reports the proportion of $\mathcal{G}^{-}$ stations in the top- $k$ selections, with the dotted line showing their candidate-set proportion of 0.625. Panel (b) reports the income-group composition at $k\!=\!3$ , $5$ , and $10$ . Solid and hatched bars represent demand-only and FairGIN selection, respectively.

Station	Group	$\hat{y}_{j}$	$S_{j}$	DO	FG
Selected by both strategies, 2 of 6 shown
Sip Ave	$\mathcal{G}^{-}$	$+0.028$	$+0.228$	✓	✓
Columbia Park	$\mathcal{G}^{-}$	$-0.063$	$+0.137$	✓	✓
Selected only by demand-only ranking
5 Corners Library	$\mathcal{G}^{+}$	$+0.007$	$+0.007$	✓	✗
Dixon Mills	$\mathcal{G}^{+}$	$-0.010$	$-0.010$	✓	✗

TABLE V : Representative deployment results on the NYC Citi Bike development split with

\alpha\!=\!0.2

. For

\mathcal{G}^{+}

S_{j}\!=\!\hat{y}_{j}

. For

\mathcal{G}^{-}

S_{j}\!=\!\hat{y}_{j}+0.2

. DO and FG denote demand-only and FairGIN selection, respectively. ✓/✗ indicates inclusion or exclusion from the top 10.

Subgroup	$n$	Income range	MAE
$\mathcal{G}^{+}$ , high income	6	$90k to$ 149k	0.343
$\mathcal{G}^{-}$ , near median	3	$60k to$ 85k	0.588
$\mathcal{G}^{-}$ , low income	2	$36k to$ 59k	0.717
$\mathcal{G}^{-}$ , data sparse †	5	Not available	1.586
$\mathcal{G}^{-}$ , overall	10	Not applicable	1.113

TABLE VI : MAE variation within the income groups for the

M\!=\!16

new stations in the NYC development split. The binary grouping masks differences in prediction error within

\mathcal{G}^{-}

. † indicates that no ACS income record is available for the corresponding census tract.

Graph Neural Networks

Graph Neural Networks learn from graph-structured data by combining information across connected nodes, edges, and larger structures.

See all 722 papers

Bikelution: Federated Gradient-Boosting for Scalable Shared Micro-Mobility Demand Forecasting
Feb 24, 2026Antonios Tziorvas, Andreas Tritsarolis, Yannis TheodoridisSpatiotemporal Forecasting Federated Learning
Department of Informatics, University of Piraeus, Piraeus, Greece
Similarity-Based Bike Station Expansion via Hybrid Denoising Autoencoders
Apr 17, 2026Oluwaleke Yusuf, M. Tsaqif Wismadi, Adil RasheedDenoising Autoencoders Geospatial Representation Learning
Department of Engineering Cybernetics Norwegian University of Science and Technology (NTNU) NO-7491 Trondheim, Norway · Department of Architecture and Planning Norwegian University of Science and Technology (NTNU) NO-7491 Trondheim, Norway
STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting
Jul 7, 2026Ye ZihaoAttention Mechanisms Spatiotemporal Forecasting
Department of System Engineering City University of Hong Kong

Feb 24, 2026J/K move · Enter open · S save

Bikelution: Federated Gradient-Boosting for Scalable Shared Micro-Mobility Demand Forecasting

Antonios Tziorvas, Andreas Tritsarolis, Yannis Theodoridis

Department of Informatics, University of Piraeus, Piraeus, Greece

Spatiotemporal Forecasting Federated Learning Privacy-Preserving ML

Open paper AskPDF arXiv

The rapid growth of dockless bike-sharing systems has generated massive spatio-temporal datasets useful for fleet allocation, congestion reduction, and sustainable mobility. Bike demand, however, depends on several external factors, making traditional time-series models insufficient. Centralized Machine Learning (CML) yields high-accuracy forecasts but raises privacy and bandwidth issues when data are distributed across edge devices. To overcome these limitations, we propose Bikelution, an efficient Federated Learning (FL) solution based on gradient-boosted trees that preserves privacy while delivering accurate mid-term demand forecasts up to six hours ahead. Experiments on three real-world BSS datasets show that Bikelution is comparable to its CML-based variant and outperforms the current state-of-the-art. The results highlight the feasibility of privacy-aware demand forecasting and outline the trade-offs between FL and CML approaches.

Abstract

Graph Neural Networks

Explore similar work

Bikelution: Federated Gradient-Boosting for Scalable Shared Micro-Mobility Demand Forecasting

Similarity-Based Bike Station Expansion via Hybrid Denoising Autoencoders

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting